請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/82848| 標題: | AGEN: 用於零次學習之屬性引導嵌入網絡 AGEN: Attribute Guided Embedding Network for Zero-Shot Learning |
| 作者: | 徐均筑 Jyun-Jhu Syu |
| 指導教授: | 陳銘憲 Ming-Syan Chen |
| 關鍵字: | 物件辨識; 零次學習; 神經網絡; 圖像特徵 Object recognition; Zero-shot learning; Neural networks; Image representations |
| 出版年 : | 2021 |
| 學位: | 碩士 |
| 摘要: | 在零次學習中,分類器可以在所有類別的語義描述的幫助下被訓練,並用以辨識來自未知類別的圖片。大部分現存的零次學習方法都學習如何將全局圖像特徵和語義標籤投影到聯合嵌入空間中。然而這樣的全局特徵無法捕捉相關語義屬性之間的外觀關係。為了建模視覺信息和語義屬性之間的關係,我們提出了屬性引導嵌入網絡(AGEN)。具體來說,AGEN 將視覺注意力機制和屬性關係建模結合到嵌入學習中,其架構主要考慮下列三者:i) 視覺注意力學習捕捉有助於識別語義屬性存在的信息區域;ii) 使用知識圖明確地考慮了屬性的關係,並利用圖卷積網絡為每個屬性學習更好的嵌入特徵;iii) 以視覺特徵和屬性嵌入作為輸入,我們設計的嵌入模塊最終輸出與預測的類最接近的語義嵌入。 我們對三個基準數據集(即CUB、SUN 和AWA2)的進行了大量實驗,並證明我們提出的方法相對現有最先進方法的具有優越性或競爭力。消融分析也驗證了我們嵌入模塊的好處。 In zero-shot learning (ZSL), a classifier is trained to recognize images from unseen classes with the aid of the semantic descriptions of all classes. Most existing works on ZSL learn to project the global image feature and semantic labels into a joint embedding space. However, such global features fail to capture the appearance relationships between correlated semantic attributes. In order to model the relations between visual information and semantic attributes, we propose the Attribute Guided Embedding Network (AGEN). Specifically, AGEN incorporates both visual attention mechanism and attribute relation modeling into embedding learning: i) Visual attention learns to capture the informative regions that help to identify the existence of semantic attributes; ii) The relations of attributes have been explicitly considered using a knowledge graph, and the GCN is utilized to learn better embedding features for each attribute; iii) Taking visual features and attribute embeddings as input, our designed embedding module outputs the final semantic embedding, to which the predicted class is the nearest one. Extensive experiments on three benchmark datasets, i.e., CUB, SUN and AWA2, demonstrate the superiority or competitiveness of our proposed method against the existing state-of-the-art approaches. The ablation analysis also validates the benefits of our embedding module. |
| URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/82848 |
| DOI: | 10.6342/NTU202102492 |
| 全文授權: | 未授權 |
| 電子全文公開日期: | 2026-09-27 |
| 顯示於系所單位: | 電機工程學系 |
文件中的檔案:
| 檔案 | 大小 | 格式 | |
|---|---|---|---|
| ntu-109-2.pdf 未授權公開取用 | 2.66 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。
